Edited by author 11-17-2005 12:27 PM
Hi Anup,

On motivation...

I think a lot of difficult high-level vision problems can be cast as finding a manifold that's embedded in some higher-dimensional search space. An example in the paper is face recognition in variable lighting. Here the approach was to cast the face as a lambertian surface, then use clustering to partition the dataset into same-surface objects. The goal here is the same as in the Sivic paper you presented -- unsupervised class discovery. The classes in this case are the individuals (same surface = same face).

You could imagine other examples. Region tracking with optical flow is one. The search space is 4D -- location and velocity for each pixel. The regions might be curves or surfaces in this search space. How can you extract these regions? The k-lines problem is essentially doing exactly this. If there were just one region, RANSAC might work well. But if there are several regions, you'd need a method such as this one.

I see...I'm not sure I understand completely what the motivation is here. Hopefully that will be cleared up tomorrow though.
Also the results do not seem to be that great (in an absolute sense, not comparatively). Is that because this particular problem has not been well studied?